Depth_Anything_3_Motifs

Reconstruction of architectural bas-relief motifs from single monocular photographs using Depth Anything 3.

Source: Thanjavur (Tanjore) temple complex, Tamil Nadu, India.
Output: Depth map + colored point cloud + fabrication-ready STL mesh — from one image, no stereo rig, no photogrammetry setup.

---

Download sample mesh (Google Drive)

Reconstructed relief mesh. 2,217,623 vertices · 4,428,998 faces. Source: single photograph, Thanjavur, 2026.

---

Research Context

Thanjavur temple architecture follows canonical Agamic rules — specifically the Manasara and Mayamata treatises — that govern the geometry of every carved motif. The resulting forms appear fluid and organic but are structurally rule-governed. This tension makes them an ideal test case for monocular depth reconstruction: can a learned depth model recover canonical geometry from a single view?

This pipeline tests that question across a dataset of 20+ motif photographs, covering typologies including Padma (lotus), Vyala (leogryph), Kirtimukha (face), and narrative frieze panels.

The downstream application is digital fabrication — specifically robotic milling and relief reproduction — where a single site photograph becomes a fabrication-ready mesh without photogrammetric multi-view capture.

---

Pipeline

Input photograph (single view)
    ↓
Depth Anything 3 inference  [DA3NESTED-GIANT-LARGE or DA3BASE]
    ↓
Depth normalization + mask generation
    ↓
Bilateral smoothing (edge-preserving)
    ↓
RGBD → Point cloud (Open3D)
    ↓
Poisson reconstruction  →  mesh_output.stl
Ball Pivoting reconstruction  →  mesh_bpa.stl
Heightfield export (z-exaggerated)  →  relief_exaggerated.stl
    ↓
Fabrication-ready STL

---

Key Parameters

These are the critical values that make the pipeline work on stone bas-relief surfaces specifically. Generic defaults produce flat or noisy results.

Parameter	Value	Why
z_exaggeration_export	24.0	Temple relief is shallow (~5–15mm). Monocular depth compresses it further. Exaggeration recovers perceptible geometry.
invert	True	DA3 returns near=low values. Relief surfaces read correctly only when inverted.
use_center_roi	True	Background stone and frame skew normalization statistics. Central ROI isolates the motif zone.
extra_erode_px	6	Halo artifacts appear at mask boundaries. Additional erosion removes them before mesh generation.
detail_boost	0.25	High-pass filter added back to normalized depth. Recovers fine surface texture without amplifying noise.
far_pct	85	Depth percentile cutoff for background masking. Retains foreground motif, discards wall context.

---

Outputs Per Motif

Each processed image produces:

• depth_map.png — normalized depth preview (8-bit)
• mesh_output.ply / .stl — Poisson reconstruction
• mesh_bpa.ply / .stl — Ball Pivoting reconstruction
• relief_exaggerated.stl — heightfield mesh, z-exaggerated, fabrication-ready

---

Environment

pip install torch torchvision
pip install numpy==2.0.2
pip install pillow<12
pip install opencv-python-headless
pip install open3d
pip install git+https://github.com/ByteDance-Seed/Depth-Anything-3.git --no-deps

NumPy version pinning is required. DA3 pulls a conflicting version without --no-deps.
Tested on Google Colab (T4 GPU). Works on local CUDA GPU with no changes.

---

Colab Notebook

Open in Google Colab

---

Dataset

Location: Thanjavur Brihadeeswarar temple complex and surrounding gopurams, Tamil Nadu.
Capture protocol:

• Camera parallel to relief surface
• Diffuse light preferred (overcast / shade)
• Scale reference in frame where possible
• No HDR processing

Motif typologies in dataset:

• Padma (lotus) — floral, radial symmetry
• Vyala (leogryph) — composite animal form
• Kirtimukha (face/mask) — frontal, high relief
• Narrative frieze panels — low relief, horizontal composition

Dataset expansion ongoing. Target: 5 typologies × minimum 4 specimens each.

---

Citation / Reference

If you use this pipeline or dataset, please cite:

Murugesan, L. (2026). Depth_Anything_3_Motifs: Monocular depth reconstruction 
of Thanjavur temple bas-relief motifs. GitHub. 
https://github.com/libishm1/Depth_Anything_3_Motifs

---

Status

• Pipeline complete and tested (single motif)
• Parameter tuning for stone bas-relief surfaces
• Batch processing script for full dataset
• Cross-motif geometry comparison
• Fabrication output validation (CNC/robotic milling)

---

Author

Libish Murugesan — researcher and lecturer in Robotics for Architecture, Riyadh.
Portfolio · LinkedIn · GitHub